moose and Q matrices

[Anton Savchenko]

Hi! 

First of all, thanks for all the new functionality in RAxML-NG 2.0! 

I have tested moose on five fungal protein MSAs, and noticed that in all cases it chose Q matrices, though not always the expected Q.yeast, but also Q. pfam and Q. insect. Is it something to worry about? Does it tell something about either my data or Q matrices?

Cheers,
Anton

[Christoph Stelz]

Hi Anton,

Thanks for trying out the new version!

The Q matrices are treated as constants from the perspective of RAxML-NG, so this is unlikely to be a bug in the model optimizer. You could rerun MOOSE with uniform rate heterogeneity (--moose-options rhas=E) to try to narrow it down.
If Q.pfam provides a lower BIC score and higher log-likelihood, then it could very well be a better fit for your dataset.

During testing, we noticed similar results where, for instance, Q.bird performed especially well on mammalian datasets. Since the Q matrices were inferred from real datasets [1], it could be the sampling of those datasets that is at play here, though we have yet to do a systematic analysis on why this happens.

Kind regards,
Christoph


[1]: Minh BQ, Dang CC, Vinh LS et al. QMaker: Fast and Accurate Method to Estimate Empirical Models of Protein Evolution. Syst Biol 2021;70(5):1046–60. https://doi.org/10.1093/sysbio/syab010.